Anomaly Detection

Detect outliers using quartile-based boundaries:

• Calculate Q1, Q3, and IQR = Q3 - Q1
• Flag values < Q1 - 1.5×IQR or > Q3 + 1.5×IQR
• Robust to non-normal distributions

Method 3: Time Series Decomposition

Identify trend and seasonal anomalies:

• Decompose series into trend, seasonal, and residual components
• Analyze residuals for unexpected deviations
• Best for identifying pattern breaks

Method 4: Moving Average Envelope

Track deviations from rolling average:

• Calculate N-day moving average and standard deviation
• Flag values outside ±2σ envelope
• Effective for trending metrics

Instructions

Step 1: Collect Historical Data

Retrieve at least 90 days of historical values for the metric being analyzed.

Step 2: Apply Statistical Tests

Run multiple anomaly detection methods and compare results.

Step 3: Calculate Anomaly Scores

Assign severity scores (1-10) based on:

• Magnitude of deviation
• Number of methods flagging the value
• Business impact of the metric
• Historical precedent

Step 4: Contextualize Findings

Check for:

• Known corporate actions (acquisitions, divestitures)
• Market events (crisis periods)
• Data quality issues (missing data, late feeds)
• System changes (calculation updates)

Step 5: Generate Alert

Produce human-readable explanation with:

• What is anomalous
• Why it's unusual (statistical basis)
• Potential causes
• Recommended actions

Output Format

{
  "anomalies_detected": <count>,
  "critical_alerts": [
    {
      "metric": "<metric_name>",
      "date": "<date>",
      "value": <number>,
      "expected_range": [<lower>, <upper>],
      "deviation_magnitude": <number>,
      "severity_score": <1-10>,
      "detection_methods": [<list>],
      "possible_causes": [<list>],
      "recommended_actions": [<list>]
    }
  ]
}

Configurable Parameters

• lookback_days: Historical window (default: 90)
• z_threshold: Z-score cutoff (default: 3.0)
• iqr_multiplier: IQR boundary multiplier (default: 1.5)
• min_severity_score: Alert threshold (default: 7)

References

See references/statistical_methods.md for detailed formulas and implementation notes.